Incident response is a sequence of evidence updates, not one model answer.
Logs, metrics and configuration can contradict an initial hypothesis. A credible agent must retain the investigation trail, distinguish observation from inference, and stop before operational change unless the action is allowed and approved.
A control room around the investigation agent.
Investigation state, hypotheses, tool calls and approval remain tied to one inspectable session.
Logs, metrics and configuration are separate tool boundaries rather than prompt-only context.
Remediation runs inside a constrained Docker execution path.
The agent cannot invent arbitrary operational commands.
Proposed remediation pauses until the operator explicitly authorizes it.
The session checks post-action evidence instead of assuming the change worked.
The hypothesis changed when the telemetry changed.
A retained synthetic run shows the agent revising its hypothesis, proposing a bounded remediation, receiving explicit approval and verifying recovery from the same investigation session.
It does not establish production SRE performance, sustained recovery, lower MTTR, broad service coverage, production security approval or Microsoft production-agent behavior.
An observable reference slice, not autonomous production operations.
The current system demonstrates investigation workflow, tool boundaries, evidence retention, approval and verification in a local/synthetic environment. Production connectors, policy ownership, tenancy, secrets management, workload isolation, rollback strategy and service-level validation remain deployment-specific work.